Conformers

ACONFL

Summary

Performance in predicting relative conformer energies of 12 C12H26, 16 C16H34 and 20 C20H42 conformers. Reference data from PNO-LCCSD(T)-F12/ AVQZ calculations.

Metrics

  1. Conformer energy error

For each complex, the the relative energy is calculated by taking the difference in energy between the given conformer and the reference (zero-energy) conformer. This is compared to the reference conformer energy, calculated in the same way.

Benchmark speed

Fast: tests are likely to take minutes to run on CPU.

Data availability

Input structures:

  • Conformational Energy Benchmark for Longer n-Alkane Chains Sebastian Ehlert, Stefan Grimme, and Andreas Hansen The Journal of Physical Chemistry A 2022 126 (22), 3521-3535 DOI: 10.1021/acs.jpca.2c02439

Reference data:

  • Same as input data

  • \(PNO-LCCSD(T)-F12/ AVQZ\) level of theory: a local, explicitly correlated coupled cluster method.

Folmsbee

Summary

Performance in predicting relative conformer energies for a set of drug-like molecules. Each molecule has a small set of conformers (3-10 per molecule) whose energies are ranked relative to the lowest-energy conformer. Reference data from DLPNO-CCSD(T) calculations.

Metrics

For each molecule, the relative energy of every conformer is taken with respect to the lowest-energy reference conformer (set to zero). The predicted energies are aligned to the same reference conformer and converted to kcal/mol before being compared against the \(DLPNO-CCSD(T)\) reference.

  1. MAE

The mean absolute error between predicted and reference relative energies, averaged over all molecules.

  1. Conformer Score

For each molecule, the per-molecule MAE and RMSE are passed through a soft threshold (MAE at 0.5 kcal/mol, RMSE at 1.5 kcal/mol) to give a value between 0 and 1, and these are averaged across all molecules. Molecules for which the model fails to produce an energy profile score 0.

Benchmark speed

Fast: minutes on GPU, and minutes to tens of minutes on CPU. The benchmark is marked fast.

Data availability

Input structures:

  • Assessing conformer energies using electronic structure and machine learning methods Dakota Folmsbee, Geoffrey Hutchison International Journal of Quantum Chemistry 2020 121 (1) e26381 DOI: 10.1002/qua.26381

Reference data:

  • Same as input data

  • \(DLPNO-CCSD(T)\) level of theory: a local coupled-cluster method.

TorsionNet500

Summary

Performance in predicting torsional energy profiles for 500 diverse organic molecular fragments. Reference data from DLPNO-CCSD(T)/CBS calculations.

Metrics

  1. RMSE of relative torsional energy profile

  2. MAE of relative torsional energy profile

For each fragment, a torsion scan samples the energy at a series of dihedral angles. Both the reference and predicted energies are mean-centered per scan, so only the shape of the profile is compared rather than absolute energy offsets. RMSE and MAE are calculated between the reference and predicted profiles for each scan, then averaged across all 500 fragments.

Computational cost

Fast: tests are likely to take minutes to run on GPU, or less than an hour on CPU for each model, since each fragment requires a single-point energy calculation across ~20 conformers in its torsion scan, repeated for all 500 fragments.

Data availability

Input structures:

  • B. K. Rai, V. Sresht, Q. Yang, R. Unwalla, M. Tu, A. M. Mathiowetz, and G. A. Bakken, TorsionNet: A Deep Neural Network to Rapidly Predict Small-Molecule Torsional Energy Profiles with the Accuracy of Quantum Mechanics, Journal of Chemical Information and Modeling 62 (2022), 785-800. PMID: 35119861.

Reference data:

  • J. L. Weber, R. D. Guha, G. Agarwal, Y. Wei, A. A. Fike, X. Xie, J. Stevenson, B. Santra, R. A. Friesner, K. Leswing, M. D. Halls, R. Abel, and L. D. Jacobson, Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties, arXiv:2505.06462 (2025).

  • DLPNO-CCSD(T)/CBS level of theory.